Many modern NLP systems rely on word embeddings, previously trained in anunsupervised manner on large corpora, as base features. Efforts to obtainembeddings for larger chunks of text, such as sentences, have however not beenso successful. Several attempts at learning unsupervised representations ofsentences have not reached satisfactory enough performance to be widelyadopted. In this paper, we show how universal sentence representations trainedusing the supervised data of the Stanford Natural Language Inference datasetscan consistently outperform unsupervised methods like SkipThought vectors on awide range of transfer tasks. Much like how computer vision uses ImageNet toobtain features, which can then be transferred to other tasks, our work tendsto indicate the suitability of natural language inference for transfer learningto other NLP tasks. Our encoder is publicly available.
展开▼
